Skip to content

Self-hosted, Self-Improving, Federated Everything App — dev management, agent orchestration, persistent mind, digital twin, second brain, and creative pipeline from a single dashboard over Tailscale to port the user into the OS

6.4 KB
Raw Blame
/**
 * Mood board → style-relevant text context (pure).
 *
 * Shared by the style synthesis run (`moodBoardStyleSynthesis.js`) and Code
 * Animation's art direction, and kept dependency-light so a caller that only
 * needs the board's text fragments doesn't import the provider stack.
 */

import { trimTo } from './textUtils.js';

// Context bounds: a board caps at 500 items, but the synthesis context must
// stay well inside a chat-completion window. Items are taken in board order
// (the user's curation order); fragments beyond either cap are dropped and
// the count is reported in the result so the UI can say so. The aggregate
// character budget is the load-bearing one — 60 items can each carry four
// 600-char fields, far past what a small local model's window fits — and it
// is a FIXED conservative budget (≈6k tokens) rather than model-aware:
// provider window metadata isn't reliably known here, and a proposal
// synthesized from the first N curated items beats a request the model
// truncates or rejects.
const CONTEXT_ITEMS_MAX = 60;
const CONTEXT_FIELD_MAX = 600;
const CONTEXT_TOTAL_CHARS_MAX = 24000;

const EXTERNAL_SOURCE = /^(?:https?:|frame:)/i;

/**
 * The caption a model may read for a board item, or `''`. A pin imported from
 * X, Pinterest or another page (its `source` is that page) — and a frame cut
 * from one — can hold someone else's text: misleading at best, a prompt
 * injection at worst. That text stays on the board for display, but only the
 * prompt the item's analysis inferred from its pixels reaches a model. Captions
 * the user or PortOS wrote (gallery pins, uploads, notes) still count.
 */
export function moodBoardItemCaption(item) {
  if (typeof item?.caption !== 'string' || !item.caption.trim()) return '';
  if (typeof item.source === 'string' && EXTERNAL_SOURCE.test(item.source)) return '';
  return item.caption.trim();
}

/**
 * Reduce a board to the style-relevant text fragments the LLM sees. An item
 * contributes only what it actually carries: a text note, a caption that did
 * not come from an imported page (`moodBoardItemCaption`), and/or a persisted
 * analysis (prompt + negative + rationale). Media items without any of those
 * contribute nothing — synthesis reads text, not pixels (analyzing
 * an item is Phase 3's explicit per-item vision step).
 */
export function collectBoardStyleContext(board) {
  const items = Array.isArray(board?.items) ? board.items : [];
  const entries = [];
  for (const it of items) {
    if (!it || typeof it !== 'object') continue;
    const entry = {};
    if (it.type === 'text' && typeof it.text === 'string' && it.text.trim()) {
      entry.note = trimTo(it.text, CONTEXT_FIELD_MAX);
    }
    const caption = moodBoardItemCaption(it);
    if (caption) entry.caption = trimTo(caption, CONTEXT_FIELD_MAX);
    if (typeof it.prompt === 'string' && it.prompt.trim()) {
      entry.prompt = trimTo(it.prompt, CONTEXT_FIELD_MAX);
    }
    const analysis = it.analysis;
    if (analysis && typeof analysis === 'object' && typeof analysis.prompt === 'string' && analysis.prompt.trim()) {
      entry.analyzedPrompt = trimTo(analysis.prompt, CONTEXT_FIELD_MAX);
      if (typeof analysis.negativePrompt === 'string' && analysis.negativePrompt.trim()) {
        entry.analyzedNegative = trimTo(analysis.negativePrompt, CONTEXT_FIELD_MAX);
      }
      if (typeof analysis.rationale === 'string' && analysis.rationale.trim()) {
        entry.analysisRationale = trimTo(analysis.rationale, CONTEXT_FIELD_MAX);
      }
    }
    if (!Object.keys(entry).length) continue;
    entries.push({ kind: it.type, ...entry });
  }
  // Over budget, keep an evenly spaced sample of the WHOLE board rather than the
  // first N — truncating in board order made the composite describe only the
  // earliest pins. Order is preserved so the curated sequence still reads through.
  const entrySize = (e) => Object.values(e).reduce((sum, v) => sum + (typeof v === 'string' ? v.length : 0), 0);
  const avgSize = entries.length ? entries.reduce((sum, e) => sum + entrySize(e), 0) / entries.length : 0;
  const budgetItems = Math.min(CONTEXT_ITEMS_MAX, avgSize ? Math.floor(CONTEXT_TOTAL_CHARS_MAX / avgSize) : CONTEXT_ITEMS_MAX);
  const candidates = entries.length <= budgetItems
    ? entries
    : Array.from({ length: budgetItems }, (_, i) => entries[Math.floor(((i + 0.5) * entries.length) / budgetItems)]);
  const fragments = [];
  let totalChars = 0;
  for (const e of candidates) {
    const size = entrySize(e);
    if (fragments.length >= CONTEXT_ITEMS_MAX || totalChars + size > CONTEXT_TOTAL_CHARS_MAX) break;
    totalChars += size;
    fragments.push(e);
  }
  const dropped = entries.length - fragments.length;
  // The board's own composite prompt is the distilled through-line. It is one
  // field, so it gets the description budget rather than the per-item cap —
  // clipping it to 600 would drop the look downstream renders are asked to match.
  const style = board?.style;
  const stylePrompt = trimTo(style?.prompt, 2000) || null;
  const styleNegative = trimTo(style?.negativePrompt, 2000) || null;
  return {
    name: trimTo(board?.name, 200) || null,
    description: trimTo(board?.description, 2000) || null,
    stylePrompt,
    styleNegative,
    items: fragments,
    droppedItems: dropped,
  };
}

/**
 * Local image references a board contributes, poster first. `resolveItem`
 * is `boardItemLocalImage` (injected so this module stays free of the board
 * store). The canonical poster is the board's style reference — the same role
 * a universe's style-probe image plays — and pinned item images follow it.
 * Duplicate filenames are dropped so a poster that was also pinned ships once.
 */
export function moodBoardImageCandidates(board, resolveItem) {
  const candidates = [];
  const seen = new Set();
  const push = (kind, filename, label) => {
    if (kind !== 'image' && kind !== 'image-ref') return;
    if (typeof filename !== 'string' || !filename.trim()) return;
    const key = `${kind}:${filename.trim()}`;
    if (seen.has(key)) return;
    seen.add(key);
    candidates.push({ kind, filename: filename.trim(), label });
  };
  push('image', board?.posterImageRef, 'Mood board poster');
  const items = Array.isArray(board?.items) ? board.items : [];
  for (const item of items) {
    const asset = typeof resolveItem === 'function' ? resolveItem(item) : null;
    if (!asset) continue;
    push(asset.kind, asset.filename, moodBoardItemCaption(item) || asset.filename);
  }
  return candidates;
}